Cookbook Audit Assistant
Review Anthropic Cookbook notebooks against a consistent rubric and produce actionable reports.
The skill requires reading style_guide.md first and separates automated checks, manual review, and report generation, so its scope is reasonably constrained; however, it invokes scripts, uv, detect-secrets, and potentially external services without documenting data flows, sensitive-notebook handling, user confirmation, isolation, or rollback, so points are deducted.
The workflow and some failure messages are explicit, and validate_notebook.py handles missing files, invalid JSON, and general exceptions; however, the key path lacks reproducible execution evidence in the skill itself, depends on uv, nbconvert, detect-secrets, and a target notebook, and has limited abnormal-input and missing-tool coverage. Static calibration therefore keeps this at or below 10.
The trigger is clear for notebook review or audit requests, and the workflow asks for a path when one is not supplied; however, non-fit boundaries, exact input/output contracts, semantic false-trigger conditions, Chinese-language support, and mainland-China network reachability are not specified, so points are deducted.
The name and description match the audit capability, with clear instructions, workflow, report format, checklist, and progressive reference material. The repository supplies an MIT license, version metadata, CI, and verified official provenance, but the skill itself lacks versioning, changelog, explicit maintenance ownership or update path, and its dependency guidance is fragmented, so points are deducted.
The skill can guide a structured Cookbook notebook audit and requests concrete recommendations with line references, giving it direct practical value; however, the output is a human-authored report, depends on the notebook, style guide, and available tools, and execution was not performed, leaving substantial judgment and possible rework. Static calibration keeps this at or below 7.
The skill points to auditable primary materials including style_guide.md, a validation script, CI workflows, and repository test dependencies, making some checks traceable; however, this review used source reading only, with no execution reproduction and no dedicated tests covering the skill's key paths, so the score cannot exceed 5.
- This assessment is source-only; validate_notebook.py, notebook conversion, detect-secrets, and tests were not executed.
- The automated commands may depend on uv, network access, and an environment not fully specified by the skill; without a target notebook, the audit cannot proceed beyond requesting its path.
- The skill does not specifically define handling of sensitive notebook data, external API calls, execution isolation, user confirmation, or recovery mechanisms.
- Repository CI and test files are shared context and do not prove that this skill itself was executed or passed tests.
What it does & when to use it
This skill audits a specified Anthropic Cookbook notebook using `style_guide.md` and its scoring rubric. It combines automated checks with manual review across narrative quality, code quality, technical accuracy, and actionability and understanding. The resulting report includes an overall score, dimension scores, key issues, prioritized recommendations, and concrete revision examples. It is intended for maintaining or improving Cookbook content, not for production code review.
It first reads style_guide.md, confirms the notebook path, and runs python3 validate_notebook.py <path> to perform technical checks and generate Markdown in tmp/. The validation also runs detect-secrets against hardcoded API keys and credentials using the repository's custom plugin and baseline. It then reviews the Markdown, scores four dimensions, and produces the specified audit report.
- An Anthropic Cookbook maintainer wants to review a Jupyter notebook's structure, code, and instructional clarity before release.
- A contributor needs to identify missing learning objectives, prerequisites, explanatory text, or conclusions in a new example.
- A technical editor wants to evaluate multiple Cookbook notebooks consistently with one rubric.
- A developer wants recommendations about notebook executability, dependency declarations, and API usage patterns.
Pros & cons
- Covers instructional narrative, code quality, technical accuracy, and user actionability.
- Includes automated notebook validation, Markdown generation, and secret scanning.
- Defines a clear scoring structure, checklist, and requirements for concrete recommendations.
- Depends on `style_guide.md`, `validate_notebook.py`, and detection scripts, while complete installation steps are undocumented.
- Automation supports the audit but does not replace human review and judgment.
- The source does not provide a test suite, supported-platform list, or sample audit report.
How to install
The source material does not provide an installation command. Use .claude/skills/cookbook-audit/SKILL.md from https://github.com/anthropics/claude-cookbooks and retain its referenced style_guide.md, validate_notebook.py, and scripts/detect-secrets/ files.
How to use
Provide the notebook path, for example: Audit path/to/notebook.ipynb. The documented workflow requires reading style_guide.md first, running python3 validate_notebook.py <path>, reviewing the generated Markdown in tmp/, and then producing the report with the four scoring dimensions and required sections.